a590003

Organization of This Report
We begin in Section 1 with a brief high-level overview of the BGV cryptosystem and some important
features of the variant that we implemented and our choice of representation, as well as an overview
of the structure of our library. Then in Sections 2, 3,4 we give a bottom-up detailed description of
all the modules in the library. We conclude in Section 5 with some examples of using this library.
1 The BGV Homomorphic Encryption Scheme
A homomorphic encryption scheme [8, 3] allows processing of encrypted data even without knowing
the secret decryption key. In this report we describe the design and implementation of a
software library that we wrote to implements the Brakerski-Gentry-Vaikuntanathan (BGV) homomorphic
encryption scheme [2]. We begin by a high-level description of the the BGV variant
that we implemented, followed by a detailed description of the various software components in our
implementation. the description in this section is mostly taken from the full version of [5].
Below we denote by [·] q the reduction-mod-q function, namely mapping an integer z ∈ Z to the
unique representative of its equivalence class modulo q in the interval (−q/2, q/2]. We use the same
notation for modular reduction of vectors, matrices, and polynomials (in coefficient representation).
Our BGV variant is defined over polynomial rings of the form A = Z[X]/Φ m (X) where m
is a parameter and Φ m (X) is the m’th cyclotomic polynomial. The “native” plaintext space for
this scheme is usually the ring A 2 = A/2A, namely binary polynomials modulo Φ m (X). (Our
implementation supports other plaintext spaces as well, but in this report we mainly describe the
case of plaintext space A 2 . See some more details in Section 2.4.) We use the Smart-Vercauteren
CTR-based encoding technique [10] to “pack” a vector of bits in a binary polynomial, so that
polynomial arithmetic in A 2 translates to entry-wise arithmetic on the packed bits.
The ciphertext space for this scheme consists of vectors over A q = A/qA, where q is an odd
modulus that evolves with the homomorphic evaluation. Specifically, the system is parametrized
by a “chain” of moduli of decreasing size, q 0 > q 1 > · · · > q L and freshly encrypted ciphertexts are
defined over R q0 . During homomorphic evaluation we keep switching to smaller and smaller moduli
until we get ciphertexts over A qL , on which we cannot compute anymore. We call ciphertexts that
are defined over A qi “level-i ciphertexts”. These level-i ciphertexts are 2-element vectors over R qi ,
i.e., ⃗c = (c 0 , c 1 ) ∈ (A qi ) 2 .
Secret keys are polynomials s ∈ A with “small” coefficients, and we view s as the second element
of the 2-vector ⃗s = (1, s). A level-i ciphertext ⃗c = (c 0 , c 1 ) encrypts a plaintext polynomial m ∈ A 2
with respect to ⃗s = (1, s) if we have the equality over A, [〈⃗c, ⃗s〉] qi = [c 0 + s · c 1 ] qi ≡ m (mod 2), and
moreover the polynomial [c 0 +s·c 1 ] qi is “small”, i.e. all its coefficients are considerably smaller than
q i . Roughly, that polynomial is considered the “noise” in the ciphertext, and its coefficients grow
as homomorphic operations are performed. We note that the crux of the noise-control technique
from [2] is that a level-i ciphertext can be publicly converted into a level-(i + 1) ciphertext (with
respect to the same secret key), and that this transformation reduces the noise in the ciphertext
roughly by a factor of q i+1 /q i .
Following [7, 4, 5], we think of the “size” of a polynomial a ∈ A the norm of its canonical
embedding. Recall that the canonical embedding of a ∈ A into C φ(m) is the φ(m)-vector of complex
numbers σ(a) = (a(τm)) j j where τ m is a complex primitive m-th root of unity (τ m = e 2πi/m ) and
the indexes j range over all of Z ∗ m. We denote the l 2 -norm of the canonical embedding of a by
1
16. Design and Implementation of a Homomorphic-Encryption Library